Page segmentation is a key task in document processing, enabling effective extraction of structured information from diverse document types. This paper presents an in-depth analysis of the method proposed by Kise et al., a bottom-up approach using area Voronoi diagrams to identify spatial relationships between document parts. Our work provides a detailed description of the method, emphasizing clarity, reproducibility, and transparency, particularly regarding aspects not fully specified in the original paper. We highlight the impact of the parameter settings and preprocessing steps on the method's performance. Through extensive testing, we demonstrate that the method can handle a wide range of layouts but exhibits notable sensitivity to specific document characteristics, especially in handling complex elements like handwritten text, lists, drop-caps, and tables.
We introduce a novel image segmentation algorithm based on the methodology of approximating solutions to backward stochastic differential equations (BSDEs). The segmentation method repeats the BSDE reconstruction process, with the parameters of these equations changing in subsequent steps. We are interested in a sequence of images driven by BSDE solutions. As the segmentation result, we define the limit of these images. By their nature, stochastic tools, particularly the Monte Carlo method, have high computational complexity. There are concerns about the running time of the proposed method, especially if we are considering a sequence of stochastic solutions. Experimental segmentation results show that it is possible to obtain results quickly and that the algorithm yields excellent results for images with intense noise.
In many applications, planar spline interpolations of scattered data on the sphere are unsatisfactory; spherical splines are desired. Wahba (1981) defined the thin-plate splines on the sphere by analogy with the polynomial splines on the circle and the thin-plate splines in Rd. The thin-plate spline fit to a scattered data set on the sphere is the solution to an empirical risk minimization problem that penalizes the infidelity of the fit to the data as well as its "wiggliness". This latter term is the square of a seminorm penalty based on the Laplace-Beltrami operator. The minimization problem is posed in a reproducing kernel Hilbert space (RKHS) of functions of finite wiggliness, whose reproducing kernel is isotropic and, due to a result by Schoenberg (1942), given by a Legendre series. A closed-form expression (in terms of the polylogarithm) for the kernel was found by Wendelberger (1982) and re-discovered by Beatson and zu Castell (2018). These closed-form expressions make not just spline interpolation but also downstream signal-processing tasks, such as cubature or resolution of inverse problems, more tractable in fields where scattered data and spherical models are common, such as remote sensing, geostatistics, motion planning, graphics, and medical imaging. In this paper, we present a tutorial on spline methods in RKHSs and show how they can be used to interpolate, smooth, and numerically integrate scattered data on the sphere and solve related inverse problems. The accompanying demo compares thin-plate spline interpolation over the sphere with thin-plate splines on an equirectangular projection and natural cubic splines on a one-dimensional latitudinal projection used in greenhouse gas monitoring. Global mean values of the interpolation surfaces are presented as well, to illustrate how this isotropic spherical kernel-which penalizes interpolant wiggliness without concern for application-specific factors like atmospheric winds-affects the computation of global averages.
In this paper, we describe an implementation of the two-phase image segmentation algorithm proposed by Goldstein, Bresson and Osher in [Geometric Applications of the Split Bregman Method: Segmentation and Surface Reconstruction, Journal of Scientific Computing, 2010]. This algorithm partitions the domain of a given 2D image into foreground and background regions, and each pixel of the image is assigned membership to one of these two regions. The underlying assumption for the segmentation model is that the pixel values of the input image can be summarized by two distinct average values, and that the region boundaries are smooth. Accordingly, the model is formulated as an energy functional whose variable is a region membership function that assigns pixels to either region, as originally proposed by Chan and Vese in [Active Contours Without Edges, IEEE Transactions on Image Processing, 2001]. This energy is the sum of image data terms in the regions and a length penalty for region boundaries. Goldstein, Bresson and Osher modify the energy of Chan-Vese so that their new energy can be minimized efficiently using the split Bregman method to produce an equivalent two-phase segmentation. We provide a detailed implementation of this method, and document its performance with several images over a range of algorithm parameters.
This article deals with arrays of real numbers which have been reduced modulo 2h into the interval [-h, h] where h > 0 is a positive real number. Such an array is said to be wrapped modulo 2h. Often, the elements of these arrays correspond to values observed at points in an image-like 2D space which are connected by a graph structure. The process of retrieving the original array from which the wrapped image originates is called unwrapping. Of course, the wrapping process is not one-to-one, and the quality of the recovered unwrapped version depends on the smoothness of the original array. The goal of unwrapping is to define a most plausible left inverse (as will be defined in a precise way) to the non-injective modulation operator mod 2h using heuristic arguments and regularity assumptions on the original signal. Following the guidelines described in [M. Constantini, A Novel Phase Unwrapping Method Based on Network Programming, IEEE Transactions on Geoscience and Remote Sensing, 1998] and [M. Constantini et al., A general formulation for redundant integration of finite differences and phase unwrapping on a sparse multidimensional domain, IEEE Transactions on Geoscience and Remote Sensing, 2012], this is made possible by correcting an approximate gradient into a global gradient using either linear programming or, in some cases, minimum-cost flow techniques to solve an L-1-norm optimization problem. Such a gradient-correcting technique can also be used in general for finding a most plausible gradient and reconstructing a signal. The online demo associated with this paper implements the aforementioned methods.
This work describes the symmetric method by Kervrann et al. for change detection. The algorithm processes a pair of images using a hypothesis testing technique with an a contrario approach. We perform a brief analysis of the results produced by the method and evaluate its quality and limitations on the Sentinel-2 OSCD dataset. Source Code The source code and documentation for this algorithm are available on the web page of this article1. Usage instructions are included in the README.txt file of the archive. The code is a symmetric version of the method from Kervrann et al.
Gaussian Splatting has emerged as a powerful technique for signal representation, especially in 3D. This paper introduces Gaussian Splatting and demonstrates its application across 1D, 2D, and 3D cases. We also discuss Gaussian Splatting in relation to Neural Radiance Fields (NeRF), highlighting the computational trade-offs and performance benefits. Through this work, we aim to bridge the gap between foundational concepts in view synthesis and advanced research, making Gaussian Splatting a more approachable and widely understood technique in the field of signal processing and computer vision. We provide code examples and detailed explanations to make the topic accessible to a broader audience, enabling readers to dive into more advanced technical papers with ease.
Neural Radiance Fields (NeRF) have emerged as a leading technology for 3D digitization, especially for their high accuracy and intricate detailing. Despite their advancements, early NeRF models struggle to handle reflections on specular surfaces effectively. To address this, alternative approaches such as Ref-NeRF and NRFF were proposed to improve fidelity in representing this physical phenomenon. This study compares these two models, providing an analysis of their effectiveness and limitations in dealing with complex specularities. We demonstrate that both methods struggle with inter-reflections and tend to model anisotropic specularities by altering the predicted surface normals.
In this work, we present an innovative multidimensional tool developed for gait evaluation and monitoring in patients with neurological disorders in routine clinical practice using Inertial Sensors, named semiogram. It has previously been published and validated by Voisard et al. [C. Innovative Multidimensional Gait Evaluation using IMU in Multiple Sclerosis: introducing the Semiogram, Frontiers in Neurology, 2023]. This tool offers a quantitative semiological analysis based on average speed and 16 other gait parameters, grouped into 7 criteria recognized in the literature: sturdiness, springiness, steadiness, stability, smoothness, synchronization, and symmetry. The provided visualization aims to facilitate easy interpretation by the clinician.
Generative artificial intelligence is increasingly being applied in diverse areas such as architecture design, music composition, or character animation. Among the generative methods, diffusion models are today the state of the art in the synthesis of high quality images with inherent diversity and realism. This paper aims to evaluate the fidelity and realism of the synthesis achieved by different architectural variations of a latent diffusion model, which is used to generate aerial images conditioned to semantic maps. As shown in the results, the diffusion model tends to correctly capture the overall semantic structure and generates realistic textures, often with a lack of fine-grained detail. Among the conditioning variations, cross-attention layers were crucial to outline the semantic segments more accurately and exploit conditional data more effectively.
This work describes a Tree Ring Detection method for complete Cross-Sections of trees (CS-TRD). The method is based on the detection, processing, and connection of edges corresponding to the tree's growth rings. The method depends on the parameters for the Canny Devernay edge detector ($\sigma$ and two thresholds), a resize factor, the number of rays, and the pith location. The first five parameters are fixed by default. The pith location can be marked manually or using an automatic pith detection algorithm. Besides the pith localization, the CS-TRD method is fully automated and achieves an F-Score of 89\% in the UruDendro dataset (of Pinus Taeda) with a mean execution time of 17 seconds and of 97\% in the Kennel dataset (of Abies Alba) with an average execution time 11 seconds.
Neural rendering methods for learning the appearance and geometry of 3D scenes have gained tremendous popularity since 2020. In this field, NeRF or Neural Radiance Fields is the bestknown methodology. Given a collection of multi-view images and their camera models, NeRF optimizes a neural network to learn the color and scene geometry that render the input images according to classical volumetric rendering techniques. NeRF operates in a self-supervised manner and provides a remarkable level of detail, but the time-consuming optimization process remains a major limitation. This paper reviews the Voxel-Accelerated NeRF (VaxNeRF), a simple acceleration strategy for NeRF proposed in 2021. VaxNeRF reduces the number of point queries required in training and inference time by considering only the region of space corresponding to the visual hull, i.e., the maximum volume compatible with the object silhouettes given by the multi-view collection. VaxNeRF requires only coarse foreground-background segmentation masks and minimal changes to the original NeRF code to improve speed by a factor of 2-8, without any performance degradation.
In this article, we examine the joint InSAR phase denoising and coherence estimation performance of the network known as-Net [Sica et al., IEEE Transactions on Geoscience and Remote Sensing, 2021]. We briefly examine the method, network architecture, training data and strategy. Then, in the experimental section, we compare the network's performance against the simple boxcar uniform filter. We verify the observations made by the authors, in particular concerning the superior denoising performance and preservation of fine details in the coherence estimation. Our experiments also indicate that an end-to-end deep learning method might bring a small improvement to the patch-based approach adopted in-Net.
The estimation of the lifetime of an industrial equipment or a patient is often based on censored data, because the event of interest is observed only for a subsample of observations. The use of the Random Forest algorithm applied to industrial data is relevant because the algorithm presents robust performances in many applications. Coupled with survival approaches, it can produce time trajectories for each subset of the feature space and thus differentiate observed objects with respect to their lifetimes. Our work aims to generalize the existing tree-based approach CART applied to left-truncated right-censored data to obtain a Random Forest algorithm. We provide a simple API to use such algorithm as well as tools to validate a temporal score against censored data.
In outdoor scenes, atmospheric absortion and scattering attenuate the radiance received by the camera and may produce haze. In 2009 He et al. proposed a simple but effective dehazing algorithm based on a hypothesis called the 'dark channel prior' (DCP). Based on this prior several other dehazing methods have been published in recent years. In this paper we review the original algorithm by He et al., together with some posterior improvements proposed by the same and other authors. We also analyze the effect of the parameters on the results and we study a variant of the method proposed by Drews et al. for the analysis of haze in underwater images.
In this article, we give a thorough description of the algorithm proposed in [H. Carrillo, M. Clement and A. Bugeau, Non-local matching of superpixel-based deep features for color transfer, VISAPP, 2022] for color transfer by relying on a robust non-local correspondence between lowlevel features at high resolution. An adaptation of this method for colorization is also described. We highlight the overall relevant results obtained with this technique for both applications and also show its limitations.
This work analyzes the BigColor method, a fully automatic colorization approach that aims to meet the challenge of providing realistic and vivid colorization for complex and diverse images in real -world scenarios. The method is a BigGAN-inspired encoder -generator network, using a spatial feature map, enabling single forward -pass colorization, supporting arbitrary input resolutions, and producing multimodal colorization results. We provide a short analysis of the method's results and highlight some limitations alongside its achievements.
This paper briefly describes and analyzes iColoriT, a hybrid colorization method based on a Vision Transformer that propagates user hints to relevant regions of a grayscale image while using color priors learned from a large image dataset. This approach gives users more control over color inference and shows a quick way to achieve results.
